Optimized search over the Gabor dictionary for note decomposition and recognition

Performing real-time note detection of multiple audio sources, such as guitar signals, is a difficult task due to the complex nature of these signals. For instance, guitar notes are close in pitch and the harmonic overtones of each note are strongly interlaced, thus preventing standard filtering techniques. In order to accomplish note detection, a harmonic matching pursuit algorithm has been used to decompose an audio signal in terms of elementary waveforms called harmonic atoms. These atoms are derived from the standard matching pursuit algorithm, and are part of an extended and overcomplete Gabor dictionary. In this paper, the search over Gabor dictionary is optimized by using signal modeling of the guitar signal in order to pre-calculate a parameter set; therefore avoiding a costly search over the extended Gabor dictionary. The parameter set defined for the proposed algorithm includes time-location, decay rate, frequency, scale and phase, which are calculated at the onset of each note played. This optimized algorithm is demonstrated through synthesized and real guitar signal examples. Considerable computational savings of this proposed algorithm over the harmonic matching pursuit algorithm are achieved.